Intent-Driven Online Privacy Budget Allocation Under Adversarial AI Attacks

Consumer IoT systems increasingly rely on intent-driven decision-making, posing new challenges for online privacy resource allocation under uncertainty and AI-enabled threats. We formulate the problem of intent-driven online privacy budget allocation, where streaming requests with semantic intent and predicted risk must be processed under a global privacy budget. We propose PAOPA, a prediction-augmented online algorithm that integrates lookahead forecasts, intent-weighted risk modulation, and dynamic constraint control via primal-dual updates. We provide theoretical guarantees on regret, robustness, and consistency, even under adversarial risk distortion. Extensive experiments on three real-world datasets show that PAOPA outperforms six intent-based baselines across noise and attack levels, achieving lower cost and tighter constraint satisfaction. Our results demonstrate the practical value of PAOPA for privacy-aware decision-making in consumer electronics.